Method for identifying abnormalities in target cells in biological samples

Abstract

The present invention relates to a computer-based implementation and device for the analysis of digital cytological slides of a biological sample, wherein the biological sample is pre-collected from a subject suspected of having bladder cancer, and the device comprises at least one input configured to receive at least one digital cytological slide obtained from the biological sample, at least one processor configured to detect target cells from at least one digital cytological image, compute a feature vector for each target cell including at least one feature computed for each target cell, form a bag of k instances for each digital cytological slide, and compute a comprehensive predictive score representing the probability of presence and / or stage of bladder cancer in the subject, and at least one output configured to provide the comprehensive predictive score.

Description

Technical Field

[0001] The present invention relates to the field of whole slide image (WSI) analysis, and more particularly to the detection of abnormalities in biological samples. In particular, the present invention relates to the detection of cell abnormalities in biological samples, preferably urine samples. More specifically, the present invention relates to a method for predicting histological results from cell slides.

Background Art

[0002] Whole slide image analysis (WSI) is a digital pathology technique that involves scanning slide glasses containing tissue samples and converting them into high-resolution digital images. These digital images can be viewed, analyzed, and processed using computer algorithms and software.

[0003] WSI enables pathologists and researchers to examine large tissue samples in detail without the limitations of conventional microscopy. This allows for quantitative analysis of tissue characteristics and can assist in the diagnosis, prognosis, and treatment of various diseases, including cancer.

[0004] As a non-invasive technique, cytopathology of urine sediment is a promising method for diagnosing bladder-related disorders or diseases. However, calculating WSI from urine slides is not an easy task because it requires recognizing and differentiating abnormalities of numerous elements, including various types of cells, cylinders, crystals, bacteria, yeast, etc.

[0005] ] Furthermore, currently used computational tools generally answer in a "yes / no" format regarding suspicion of cancer. [[ID=]25]

[0006] In bladder cancer, there is a high interest in differentiating and sorting WSI of urine samples in order to determine the presence or absence of abnormalities in urothelial cells and further determine whether these abnormalities lead to suspicion of low-grade bladder cancer or high-grade cancer.

[0007] Low-grade bladder cancer is a type of non-invasive bladder cancer that grows slowly and is unlikely to spread to other parts of the body. It is often minimally invasive and, while it can recur, is unlikely to progress to muscle-invasive disease. Low-grade bladder cancer is usually treated by removing the tumor via transurethral resection, followed by close monitoring and surveillance to detect recurrence.

[0008] On the other hand, high-grade bladder cancer is a rapidly growing, more invasive, and aggressive form of bladder cancer that carries a higher risk of spreading to other parts of the body, especially if left untreated. It is more likely to recur and progress to muscle-invasive disease, which can be life-threatening if not treated promptly.

[0009] Bladder cancer grading is based on the appearance of cancer cells under a microscope. Cancer cells are classified as low-grade or high-grade based on their level of abnormality. High-grade cells are more abnormal and invasive. Determining the grade of bladder cancer is crucial for the medical team to design a treatment strategy for the patient. [Overview of the project] [Problems that the invention aims to solve]

[0010] The present invention aims to help medical teams obtain quick and accurate information regarding the next steps they should take. In particular, the present invention helps predict the results of cystoscopy in patients and helps monitor patients in the most optimal way. [Means for solving the problem]

[0011] Therefore, the present invention relates to a device for the analysis of digital cytological slides of biological samples. The biological samples are pre-collected from subjects suspected of having bladder cancer, and the device is • At least one input configured to receive at least one digital cytology slide obtained from the digitization of at least one whole slide image (WSI) of a biological sample, • At least one processor, It is equipped with the following: The processor is • Detect target cells from at least one digital cytology image, For each target cell, calculate a feature vector that includes at least one feature calculated for each target cell. • Each digital cytology slide forms a bag of k instances, where the k instances are selected as multiple target cells that are multiple high-dysplasia cells among the target cells detected for each digital cytology slide, based on at least one feature of the feature vector. • Calculate a comprehensive predictive score representing the probability of presence and / or stage of the target bladder cancer. The comprehensive predictive score is obtained from a combination of single predictive scores obtained from at least two multi-instance learning methods configured to receive at least one formed bag as input. The device is configured to provide a comprehensive predictive score.

[0012] Advantageously, the device of the present invention is based on pre-selecting the most atypical cells and predicting the diagnosis by an ensemble of multi-instance learners. In addition to the computational significance of selecting the most atypical target cells, this method is advantageous in that the "most positive" instances contain sufficient information in both positive and negative slides. In the case of positive slides, healthy cells tend to be excluded from analysis, which increases attention to the atypical cells of interest. On the other hand, in the case of negative slides, slightly atypical but healthy target cells are given higher importance, acting as hardmining that can help reduce false positives.

[0013] Advantageously, according to other aspects of the present invention, the device includes one or more of the features described in the following embodiments, either individually or in any combination.

[0014] According to one embodiment, at least one digitized cytological image is obtained from at least one WSI colored with Papanicolaou. Advantageously, this embodiment allows for obtaining isolated and stained foreground objects that are easily distinguishable from the background.

[0015] According to one embodiment, the detection of at least one target cell is performed using a trained Res-Net model. This model is configured to receive at least one preprocessed WSI portion as input and to identify and classify, as output, objects present in the at least one preprocessed WSI portion as belonging to at least one class of target cells.

[0016] In one example, the ResNet-18 model is trained to classify all foreground object crops into five classes, including basal urothelial cells (BUCs), superficial urothelial cells, aggregates (at least two contact objects), multinucleated neutrophils, and others, based on more than 25,000 annotated crops. In this invention, the target cells are basal urothelial cells (BUCs), and other classified objects may be discarded.

[0017] According to one embodiment, at least one feature included in the feature vector is selected from the nucleocytoplasmic ratio (NCR), nuclear intensity, nuclear intensity standard deviation, Haralick energy, entropy, uniformity, convex hull ratio, and / or nuclear roundness.

[0018] According to one embodiment, at least one feature is the nuclear-cytoplasmic ratio (NCR), and the selection of k target cells is as follows: • Sort the target cells in descending order of their nucleus-cytoplasm ratio, • Select the first k target cells that represent highly dysplastic basal urothelial cells, Includes.

[0019] Advantageously, the present embodiment focuses on the nuclear cytoplasmic ratio (NCR), which is the most important criterion for evaluating the atypia of BUC according to the Paris System.

[0020] According to one embodiment, if the slide contains less than k BUCs, all BUCs are selected and the bag is zero-padded to keep the bag size constant.

[0021] According to one embodiment, the calculation of the feature vector includes pre-segmentation of at least one detected target cell and its nucleus using a convolutional machine learning model, such as a U-Net model.

[0022] According to one embodiment, the convolutional machine learning model is pre-trained using random data augmentation so that the convolutional machine learning model is robust to potential staining variations.

[0023] According to one embodiment, k is a predefined constant natural number, which enables the improvement of the consistency of the bag size.

[0024] According to one embodiment, at least two multi-instance learning models are embedding-based multi-instance learning models that use a pooling operator, which is a weighted mean operator.

[0025] According to one embodiment, at least two multi-instance learning models are trained using bootstrap aggregation.

[0026] According to one embodiment, bladder cancer is urothelial carcinoma.

[0027] According to one embodiment, the biological sample is a urine sample.

[0028] According to one embodiment, the target cell is a basal urothelial cell.

[0029] According to another preferred embodiment of the present invention, the device includes one or more of the features described in the following embodiments, either individually or in any possible combination.

[0030] In addition, this disclosure relates to a computer program which includes software code adapted to perform, when the program is executed by a processor, a method for analyzing digital cytological slides of biological specimens or a method for training compliant with either of the above execution modes.

[0031] This disclosure further relates to a computer-readable non-temporary program storage device that tangibly embodies a program of computer-executable instructions for performing a method or training method for analyzing digital cytological slides of biological specimens, in accordance with this disclosure.

[0032] Such non-temporary program storage devices may be, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor devices, or any suitable combination thereof. Portable computer diskettes, hard disks, ROMs, EPROMs (erasable programmable ROMs), or flash memory, and portable CD-ROMs (compact disk ROMs) provide more specific examples, but should be understood, as will be readily apparent to those skilled in the art, these are merely illustrative examples and not an exhaustive list. [Brief explanation of the drawing]

[0033] [Figure 1] This is a general scheme illustrating an overview of one embodiment of the present invention. [Figure 2] (a) is a diagram showing an example architecture of the feature extractor f used, and (b) is a diagram showing an example architecture of the classifier g used. [Figure 3](a) is a figure showing the average attention score of all positive test slides for instances ordered in ascending order of NCR, and (b) is a figure showing the average attention score of all negative test slides for instances ordered in ascending order of NCR. In both cases, cells with higher NCR tend to have higher attention scores. [Figure 4] This figure illustrates the ABMIL performance evaluated by cross-validation over k instances selected for each bag. [Modes for carrying out the invention]

[0034] [Definition] In this invention, the following terms have the following meanings.

[0035] The terms “adapted” and “configured” are used in this disclosure to broadly encompass the initial configuration of the Device, subsequent adaptations or complements, or any combination thereof, whether achieved by apparatus or software means (including firmware).

[0036] The term “processor” should not be interpreted as being limited to hardware capable of running software, but generally refers to a processing device that may include, for example, a computer, microprocessor, integrated circuit, or programmable logic device (PLD). Furthermore, a processor may encompass one or more image processing units (GPUs), whether or not they are used for computer graphics and image processing or other functions. In addition, instructions and / or data that enable the execution of related functions and / or resulting functions may be stored in any processor-readable medium, such as integrated circuits, hard disks, optical discs such as CDs (Compact Discs) and DVDs (Digital Multipurpose Discs), RAM (Random Access Memory), or ROM (Read-Only Memory). Instructions may, in particular, be stored in hardware, software, firmware, or any combination thereof.

[0037] This description illustrates the principles of the disclosure. Therefore, it will be understood that a person skilled in the art can devise various configurations that embody the principles of the disclosure, even if not explicitly described or illustrated herein.

[0038] All examples and conditions described herein are intended for educational purposes to help readers understand the principles of this disclosure and the concepts introduced by the inventors to advance the art, and should be construed as not being limited to such specifically described examples and conditions.

[0039] Furthermore, all descriptions herein describing the principles, aspects, and embodiments of this disclosure, as well as specific examples thereof, are intended to encompass both structural and functional equivalents. Moreover, such equivalents are intended to include both currently known equivalents and equivalents to be developed in the future (i.e., any developed elements that perform the same function regardless of their structure).

[0040] Accordingly, it will be understood by those skilled in the art that, for example, the block diagrams presented herein may represent conceptual diagrams of exemplary circuits embodying the principles of this disclosure. Similarly, it will be understood that any flowcharts and flow diagrams, etc., which are substantially represented in a computer-readable medium, may represent various processes that can be performed by a computer or processor, whether or not a computer or processor is explicitly indicated.

[0041] The functions of the various elements shown in the diagram may be provided using dedicated hardware, as well as hardware capable of running software in conjunction with appropriate software. When provided by a processor, the functions may be provided by a single dedicated processor, a single shared processor, or multiple individual processors, some of which may be shared.

[0042] It should be understood that the elements shown in the figure may be implemented in various forms of hardware, software, or combinations thereof. Preferably, these elements are implemented as a combination of hardware and software on one or more appropriately programmed general-purpose devices, which may include a processor, memory, and input / output interfaces.

[0043] This invention uses multiple instance (MI) learning (also known as multi-instance learning) as an alternative to conventional supervised learning models. In MI learning methods, learning examples are represented by a bag of instances (i.e., a multiset) instead of a single feature vector.

[0044] In typical machine learning problems such as image classification, it is assumed that images clearly represent a category (class). However, in many real-world applications, multiple instances are observed, and only a general description of the category is given. This scenario is called multiple instance learning (MIL) or learning from poorly annotated data. The problem of poorly annotated data is particularly evident in medical imaging (e.g., computational pathology, mammography, or CT lung screening) where images are typically described by a single label (benign / malignant) or where a region of interest (ROI) is roughly given. MIL deals with bags of instances to which a single class label is assigned. Therefore, the main objective of MIL is to train a model to predict the bag label.

[0045] MI learning methods include, non-restrictively, instance-based MIL such as MIWrapper, e.g., ABMIL, TRANSMIL, or embedding-based MIL, also known as DSMIL or dual-stream MIL, such as wrapper algorithms like MILES and SimpleMI.

[0046] The instance-based MIL method first predicts the label yi for each instance i, and then attempts to estimate a comprehensive label from the resulting set of yi|i∈[1,k].

[0047] MIWrapper performs propositionalization by assigning bag-level class labels to instances and weighting the instances so that each bag has the same overall weight. A single-instance model is built on the resulting dataset, and bag-level predictions are made by averaging the predicted probabilities of the instances within each bag.

[0048] Embedding-based methods first construct a representation (embedding) of the bag from which a comprehensive label is predicted. Each MIL method consists of a feature extractor f(x): Rd → Rd', a classifier g(h): Rd' → R, and a permutation-invariant pooling operator σ.

[0049] The ABMIL method used in this invention is an attention-based MIL that performs an embedding where the pooling operator is a weighted average operator, σ°f(X)=PN i=1 aihi, where hi=f(xi) and ai is the attention score associated with instance i, learned by the model from hi. ai=exp{wT}.

[0050] TransMIL is an embedding-based method that relies on a self-attention layer from the transformer to calculate an instance's attention score by referencing all other instances in the bag. Unlike previous methods, the importance of an instance is then determined using pairwise comparisons with all other instances in the bag.

[0051] Multiple Instance Learning with Embedding Instance Selection (MILES) is a recent MI learning technique that transforms MI data into a propositional form, applying a one-norm support vector machine (SVM) classifier. MILES embeds bags into a single-instance feature space. MILES uses a symmetric assumption that multiple target points are allowed, each of which can be associated with either a positive or negative bag.

[0052] SimpleMI performs proposition by averaging the attribute values ​​of instances within each bag and adding the bag's class label to the resulting features.

[0053] Interestingly, the SimpleMI method generates only one instance for each training bag without increasing the dimensionality of the feature space. For example, MIWrapper generates one instance for all instances in all bags, leaving the dimensionality of the feature space unchanged. In contrast, if MILES similarly generates one instance per training bag, the dimensionality of the feature space will almost always be much higher, as the number of attributes is equal to the total number of instances in the training bag.

[0054] The dual-stream MIL method combines both instance-based and embedding-based techniques. An instance-based max pooling branch is used to predict the malignancy score for each instance, with the instance with the highest score being kept as critical. A second embedding-based branch uses a weighted average operator as σ, and the attention score ai is calculated as the distance to the critical instance. The prediction is the average of the outputs of each branch.

[0055] In one embodiment, all methods used in the present invention include instance selection and ensemble or bootstrap aggregation techniques.

[0056] In one embodiment, the method of the present invention may employ an ensemble because the MIL method is a weakly supervised method, and therefore the resulting predictions may vary and are usually evaluated multiple times per experiment. To mitigate this variability, an ensemble is used during inference, training N individual MIL models on N independent subsets of the training data and averaging the raw predictions.

[0057] In one embodiment, the method of the present invention may perform bootstrap aggregation, also known as bagging, which is an ensemble machine learning technique used to improve the stability and accuracy of machine learning models. It works by combining multiple models built on different samples of the training data to produce a single, more robust model. The idea behind bagging is to reduce the variance of a model that can occur when a model overfits the training data. By training multiple models on different samples of the training data, bagging helps to capture the underlying structure of the data more accurately. In bootstrap aggregation, the training data is randomly sampled by substitution to create multiple training sets. A separate model is then trained on each of these training sets. The final prediction is made by combining the predictions from each of the models. This can be done by taking a majority vote in the case of a classification problem, or by averaging the predictions in the case of a regression problem. [Examples]

[0058] The present invention is further illustrated by the following embodiments.

[0059] [Examples] A standard urine cytology preparation protocol has been standardized for VisioCyt 1 MD. · Chemical fixation: Hologic Cytolyt · Centrifugal separation • Cell resuspension: Hologic Preservation • Cell deposition: Hologic Thinprep (10mm circular spot) • Staining: Papanicolaou staining • Slide mount: Sakura film Slides were digitized along 3 to 5 focal planes (×40, 0.23 μm / pixel) using Hamamatsu NanoZoomer® S360. Slides from 731 patients (268 negative, 233 LGUC, 230 HGUC) were used for training, and the remaining patients (141 negative, 76 LGUC, 75 HGUC) were stored for the test set.

[0060] 1. Detection of target cells in WSI: The thin-layer urine cytology WSIs under consideration were stained with Papanicolaou stain, which made isolated and stained foreground objects easily distinguishable from the background. Foreground objects were defined as connected components detected by Otsu's automated threshold, having regions ranging from 1200 to 90,000 pixels at full resolution. Foreground objects were detected from the central focal plane of the slide, and then crops containing each foreground object were extracted from the WSI at the sharpest focal plane, calculated using the variance of the Laplacian operator. A ResNet-18 model was trained to classify all foreground object crops into five classes based on over 25,000 annotated crops: basal urothelial cells (BUCs), superficial urothelial cells, aggregates (at least two contact objects), multinucleated neutrophils, and others.

[0061] In this example, only the BUC was considered, and all other objects were discarded.

[0062] Expert Feature Vector Calculation: Handcrafted feature vectors were calculated for each BUC. These features are called experts because they are inspired by the features described in the Paris System. These features include four nuclear texture features, including nuclear intensity standard deviation and Haralick's energy, entropy, and homogeneity, as well as two nuclear morphology features, including convex hull ratio and nuclear roundness, and nuclear cytoplasmic ratio (NCR) and nuclear intensity. The calculation of these expert features relies on accurate segmentation of cells and their nuclei. To do this, a U-Net model was trained on over 11,000 cells using annotated nuclear masks. In this example, random data augmentation, including color and intensity jittering, inversion, rotation, and cropping, was used during training to ensure that the nuclear masks predicted by the U-Net model are robust to potential staining variability.

[0063] 3. Comprehensive Diagnostic Prediction: Multiple Instance Learning (MIL) is a weakly supervised learning method designed to predict labels from a bag of instances and is well-suited to the problem of predicting diagnostic scores from a set of BUCs. In this study, MIL was used for binary classification (negative or positive for urothelial carcinoma). That is, the model aimed to predict a diagnostic score y∈[0,1] from a bag of N instances X={x1,x2,…,xk}. The label yi for each instance i∈[1,k] was unavailable. The MIL framework assumed that a bag was positive if it contained at least one positive instance.

[0064] 3. Bag Construction: While the MIL method can handle bags with a variable number of instances, in this work, the number of instances k in the bag was fixed with respect to the bag size. The inventors selected the k most heteromorphic BUCs in each digital slide to construct a bag. According to the Paris System, NCR is the most important criterion for evaluating the heteromorphism of BUCs; therefore, all BUCs in WSI were sorted in descending order of NCR, and the first k BUCs were included in the bag. In addition to computational interest, this method was motivated by the fact that sufficient information is contained in the "most positive" instances for both positive and negative slides. For positive slides, healthy cells tend to be excluded from analysis, which increased attention to heteromorphic BUCs. On the other hand, for negative slides, slightly heteromorphic but healthy BUCs were given higher importance, which acted as hard mining that could help reduce false positives. If a slide contained fewer than k BUCs, all BUCs were selected and zero-padded into the bag to keep the bag size constant. After evaluation using cross-validation, the value k=300 was used as follows.

[0065] 4. MIL method: In this embodiment, the method is a pooling operator and a weighted average operator.

number

number

number

number

[0066]

number

[0067] Figure 2 shows the feature extractor f and the classifier g.

[0068] Ensemble MIL: Because the MIL method is a weakly supervised method, the predictions obtained can vary and are usually evaluated by repeating the process multiple times per experiment. To mitigate this variability, an ensemble was used during inference, training N individual MIL models on N independent subsets of the training data and averaging the raw predictions. Each subset consisted of (N-1) / N of the entire training dataset and was calculated similarly to N-fold cross-validation. This means that each sample is within N-1 of the N training subsets.

[0069]

number

number

[0070] [result] Diagnostic Performance: Diagnostic performance was evaluated for the test set patients using precision, sensitivity, specificity, and ROC-AUC, with the experiment repeated five times and averaged, and compared to expert diagnoses. According to the Paris System, slides labeled C2 by experts were considered negative, and slides labeled C4, C5, or C6 were considered positive. Slides were rejected and labeled C1 when they did not contain basal urothelial cells to confirm the diagnosis. Furthermore, the Paris System includes a C3 class corresponding to atypical slides that contain some atypical cells but not enough to be classified as suspected HGUC. Therefore, this C3 label can be considered an indeterminate class. In the proposed method, slides were automatically rejected if fewer than 10 basal urothelial cells were detected, and slides were labeled indeterminate when the diagnostic prediction was 0.45–0.55. Thus, the proportion of rejected and indeterminate slides returned by our method matched the proportion of C1 and C3 slides labeled by experts. Table 1 shows the calculated performance for slides that were neither rejected nor unclear.

[0071] [Table 1]

[0072] This experiment demonstrates that the proposed method can diagnose bladder cancer with higher accuracy and sensitivity but lower specificity than experts. However, it should be noted that in the VisioCyt® clinical trial, negative patients could not have positive cytology (C4, C5, or C6), and the specificity of experts is 100% by composition, which is certainly an overestimation.

[0073] Model predictions were analyzed for various patient subgroups, including negative, LGUC, and HGUC. The results are shown in Table 2. Most errors occurred in LGUC patients, which are known to be difficult to detect using urine cytology.

[0074] [Table 2]

[0075] [Table 3]

[0076] Attention Score Interpretation: Analyzing ABMIL's attention score for a single slide can visualize the cells that contribute most to bag embedding and, consequently, to prediction. Furthermore, analyzing the averaged attention scores across multiple slides, taking into account any expert features, can help understand how the ABMIL model uses such features for prediction. Figure 3 shows the average attention scores for bags of instances ordered by NCR in ascending order.

[0077] On average, basal urothelial cells with higher NCR tended to have higher attention scores in the ABMIL method for both positive and negative slides. A higher NCR is the most important dysplasia criterion in the Paris System. This suggests that the attention mechanism in ABMIL does indeed tend to give higher importance to the most dysplasia cells in a slide for predicting malignancy scores. This observation also reinforces the intuition that selecting the most dysplasia cells based on expert knowledge to form bags is relevant to the MIL model.

[0078] [[Comparative and Ablation Studies]] MIL method: The ABMIL method was compared with two recent MIL methods that share a common baseline and interesting characteristics.

[0079] Baseline: The baseline methods considered depend on whether σ is the mean pooling operator or the maximum pooling operator. Depending on whether the pooling operator σ is applied before or after the classifier g, these were rejected as instance-based or embedding-based.

[0080] DSMIL: The Dual-Stream MIL method combines both instance-based and embedding-based techniques. An instance-based max pooling branch is used to predict the malignancy score of each instance, and the instance with the highest score is kept as a critical instance. A second embedding-based branch uses a weighted average operator as σ and an attention score a i This was calculated as the distance to important instances. The prediction is the average of each branch output.

[0081] TransMIL: The TransMIL method is an embedding-based method that relies on a self-attention layer from a transformer to compute an instance's attention score by considering all other instances in the bag. Unlike the method described above, the importance of an instance was then determined using pairwise comparison with all other instances in the bag. All methods used the proposed instance selection and ensemble techniques and shared similar f and g (see Figure 2).

[0082] Bag preparation: To study the effect of the proposed instance selection method, the performance of the ABMIL model was evaluated using bags containing the top 300 atypical cells and bags containing 300 randomly selected cells in the test set. The results are shown in Table 4.

[0083] [Table 4]

[0084] Figure 4 shows the effect of the number of instances k selected within the bag on ABMIL model performance, as evaluated by cross-validation, for k ∈ {100, ..., 500}. It can be seen that the number of instances k has little impact on diagnostic performance.

[0085] Ensemble: The effect of the number N of models used in the ensemble was studied compared to the special case N=1, where a single model is trained on the complete training dataset without an ensemble (see Table 5). At low N, the ensemble model was trained on a small training subset and showed lower performance. At higher N, performance was somewhat more stable. To study the effect of ensembles in the case of N=1, five models obtained by repeating the experiment were ensembled as described in the paragraph on ensemble MIL. This experiment shows that ensembles lead to performance improvements when ensemble models are trained on independent training subsets.

[0086] [Table 5]

[0087] This example demonstrates the efficiency of the developed method for predicting diagnostic status derived from the gold standard ground truth in a clinical trial dataset.

Claims

1. A device for analyzing digital cytological slides of biological samples, wherein the biological samples are pre-collected from subjects suspected of having bladder cancer, and the device is At least one input configured to receive at least one digital cytology slide obtained from the digitization of at least one whole slide image (WSI) of the biological sample, At least one processor, The target cells are detected from at least one of the aforementioned digital cytology images. For each of the target cells, a feature vector is calculated that includes at least one feature calculated for each of the target cells. For each digital cytology slide received, a bag of k instances is formed, wherein the k instances are selected as a plurality of target cells that are a plurality of highly dysplastic cells among the target cells detected for each of the digital cytology slides, based on the at least one feature of the feature vector. A comprehensive predictive score representing the probability of presence and / or stage of bladder cancer in the subject is calculated, and the comprehensive predictive score is configured to be obtained from a combination of single predictive scores obtained from at least two separate multi-instance learning methods configured to receive at least one formed bag as input. At least one processor, At least one output configured to provide the aforementioned comprehensive prediction score, A device equipped with the following features.

2. The device according to claim 1, wherein at least one of the digital cytological images is obtained from at least one of the WSI colored with Papanicolaou.

3. The device according to claim 1 or 2, wherein the detection of at least one target cell is performed using a trained Res-Net model configured to receive at least one preprocessed portion of the WSI as input and to identify and classify objects present in the at least one preprocessed portion of the WSI as belonging to at least one class of the target cell as output.

4. The device according to any one of claims 1 to 3, wherein the at least one feature included in the feature vector is selected from the nucleocytoplasmic ratio (NCR), nuclear intensity, nuclear intensity standard deviation, Haralick energy, entropy, uniformity, convex hull ratio, and / or nuclear roundness.

5. The at least one of the features is the nuclear-cytoplasmic ratio (NCR), and the selection of k of the target cells is, The aforementioned target cells are sorted in descending order of their nucleus-cytoplasm ratio, Selecting the first k target cells that represent highly dysplastic basal urothelial cells, The device according to claim 4, including the device described in claim 4.

6. The device according to any one of claims 1 to 5, wherein the calculation of the feature vector includes preliminary segmentation of the detected at least one target cell and its nucleus using a convolutional machine learning model, such as a U-Net model.

7. The device according to claim 6, wherein the convolutional machine learning model is pre-trained using random data augmentation.

8. The device according to any one of claims 1 to 7, wherein k is a predefined constant natural number.

9. The device according to any one of claims 1 to 8, wherein at least two multi-instance learning models are embedding-based multi-instance learning models that use a pooling operator which is a weighted average operator.

10. The device according to any one of claims 1 to 9, wherein the at least two multi-instance learning models are trained using bootstrap aggregation.

11. The device according to any one of claims 1 to 10, wherein the bladder cancer is urothelial carcinoma.

12. The device according to any one of claims 1 to 11, wherein the biological sample is a urine sample.

13. A computer-aided method for the analysis of digital cytological slides of a biological sample, wherein the biological sample is previously collected from a subject suspected of having bladder cancer. The steps include receiving at least one digital cytology slide obtained from the digitization of at least one whole slide image (WSI) of the biological sample, The steps include detecting target cells from at least one of the digital cytology images, For each of the target cells, a feature vector is calculated that includes at least one feature calculated for each of the target cells. A step of forming a bag of k instances for each digital cytology slide, wherein the k instances are selected as a plurality of target cells that are a plurality of highly dysplastic cells among the target cells detected for each of the digital cytology slides, based on the at least one feature of the feature vector. A step of calculating a comprehensive predictive score representing the probability of the presence and / or stage of the bladder cancer in the subject, wherein the comprehensive predictive score is obtained from a combination of single predictive scores obtained from at least two multi-instance learning methods configured to receive at least one formed bag as input, A computer-based method, including one that includes such methods.

14. A computer program product comprising instructions, wherein, when the program is executed by a computer, the instructions cause the computer to execute the computer implementation method described in claim 13.

15. A computer-readable storage medium containing instructions, wherein, when executed by a computer, the instructions cause the computer to execute the computer-implemented method described in claim 13.

Images